Pollen dataset Sebastian Seurig, Institute for Medical Informatics and Biometry, Carl Gustav Carus Faculty of Medicine, TU Dresden (https://orcid.org/0000-0001-6511-0102) Peter Steinbach, Helmholtz AI, Artificial Intelligence Cooperation Unit, Helmholtz-Zentrum Dresden-Rossendorf (https://orcid.org/0000-0002-4974-230X) Nico Scherf, Institute for Medical Informatics and Biometry, Carl Gustav Carus Faculty of Medicine, TU Dresden and Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig (https://orcid.org/0000-0003-4003-9121) Ingo Röder, Institute for Medical Informatics and Biometry, Carl Gustav Carus Faculty of Medicine, TU Dresden This artificial dataset was created for testing machine learning models on segmentation or bounding box regression and classification tasks. It contains 80.000 (1280x1280 RGB) images of airborne pollen of different sizes within their natural range: 2x10.000 images each Chenopodium bonus-henricus and Corylus colurna making up the medium-sized pollen 10.000 images Urtica dioica making up the smaller pollen 10.000 images Secale cereale making up the larger pollen 4x10.000 images each containing artifacts such as dust or bursted pollen and a mix of pollen in the following ratios: - 'equal_split' 25% each class - 'smaller_pollen_split' 70% Urtica dioica, * - 'middle_pollen_split' 40% Corylus colurna, 10% Chenopodium bonus-henricus, * - 'bigger_pollen_split' 70% Secale cereale, * - * the rest of the pollen split equally between the remaining classes Labels consist of : monochrome masks of each pollen slide (ignoring artifacts) for segmentation x and y-coordinates of the bounding boxes containing all pixels of each of the pollen for regression class names for each labeled pollen. For further questions and suggestions, please do not hesitate to contact the authors. Extensive code and configuration settings for the generator used can be found at https://github.com/seurig/slide-generator. All raw images used to generate this dataset were taken from https://pollen.tstebler.ch/.